Junhao Chen, Xiang Li, Xiaojun Ye, Chao Li, Zhaoxin Fan, Hao Zhao

Based on the LMM we developed Idea23D, a multimodal iterative self-refinement system that enhances any T2I model for automatic 3D model design and generation, enabling various new image creation functionalities togther with better visual qualities while understanding high level multimodal inputs.

• LMM: prepare the OpenAI GPT-4V API key, or use another open source LMM (e.g., LLaVA, Qwen-VL).

• Text-2-Image model: use SD-XL, DALL·E, or Deepfloyd IF.

• Image-2-3D model: use TripoSR, Zero123, Stable Zero123, or Wonder3D

❗If different modules are used, install the corresponding dependency packages.

The code we have given to run locally uses llava-1.6, SD-XL and TripoSR. so requirements-local.txt is following that.

It's driven by GPT4V, SD-XL(replicate), and TripoSR if you're using colab for testing, it uses this requirements-colab.txt.

pip install -r requirements-local.txt

Then change the path to your path in the "Initialize LMM, T2I, I23D" section of ipynb.

https://huggingface.co/llava-hf/llava-v1.6-34b-hf
https://huggingface.co/stabilityai/TripoSR
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0

This section in ipynb:

# init LMM,T2I,I23D
log('loading lmm...')

# lmm = lmm_gpt4v('sk-your open ai key')
lmm = lmm_llava_34b(model_path = "path_to_your/llava-v1.6-34b-hf", gpuid = 5)
# lmm = lmm_llava_7b(model_path = "path_to_your/llava-v1.6-mistral-7b-hf", gpuid = 2)

log('loading t2i...')
# t2i = text2img_sdxl_replicate(replicate_key='r8_ZCtxxxxxxxxxxxxxx')
t2i = text2img_sdxl(sdxl_base_path='path_to_your/stable-diffusion-xl-base-1.0', 
                    sdxl_refiner_path='path_to_your/stable-diffusion-xl-refiner-1.0', 
                    gpuid=2)

log('loading i23d...')
i23d = img23d_TripoSR(model_path = 'path_to_your/TripoSR' ,gpuid=2)
log('loading finish.')

open Idea23D/idea23d_pipeline.ipynb, Explore freely in the notebook ~

Using GPT4V, SD-XL or DALL·E, TripoSR as LMM was able to get the best results so far. The effects in the paper were obtained using Zero123, so they are inferior compared to TripoSR.

If you don't have access to GPT4V you can use Qwen-VL or LLaVA, if you use LLaVA it is recommended to use the llava-v1.6-34b model. Although we gave a pipeline built with llava-v1.6-mistral-7b, it works poorly, while llava-v1.6-34b can correctly fulfill user commands.

✅1. Release offline version of Idea23D implementation (llava-1.6-34b, SD-XL, TripoSR)

✅2. Release online running version of Idea23D implementation (GPT4-V, SD-XL, TripoSR)

🔘3. Release complete rendering script with 3d model input support. we have encountered an issue where we can't render vertex shaded objs, only objs with texture maps. if you know how to handle this please contact us. Will release the rendering part of the code after resolving this issue.

🔘4. Components supported by release: Qwen-VL, Zero123, DALL-E, Wonder3D, Stable Zero123, Deepfloyd IF. The release date for the complete set of all components will be delayed due to ongoing follow-up work.

@article{chen2024idea23d,
  title={Idea-2-3D: Collaborative LMM Agents Enable 3D Model Generation from Interleaved Multimodal Inputs}, 
  author={Junhao Chen and Xiang Li and Xiaojun Ye and Chao Li and Zhaoxin Fan and Hao Zhao},
  year={2024},
  eprint={2404.04363},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

We have intensively borrow codes from the following repositories. Many thanks to the authors for sharing their codes.

Qwen-VL, LLaVA, TripoSR, Zero123, Stable Zero123, Wonder3D, SD-XL, Deepfloyd IF